PANGLOSS: Knowledge-Based Machine Translation

What it is about

PANGLOSS is a project report on a knowledge-based interlingual system that combines symbolic and statistical methods for newspaper text in several languages. Its status list gives concrete sizes: an ontology of about 50,000 items, a 130,000-entry word list for the analyser, a bilingual Japanese-English dictionary of about 70,000 entries, and an English generation lexicon of about 70,000 items. One statistical component learns about 200,000 rules that decide where the articles "the" and "a" belong in English text produced from Japanese, which has no articles. The preliminary figure is that Spanish-to-English translators worked about 40 percent faster with the system than by hand. The report gives no method detail and no published score.

Ideas

• I1. Learn function-word insertion as counted rules. For the encoder, learn when a determiner, or a word such as "that", is needed from the concept, its tag and its neighbours, the same back-off order as profile.rs. Measure precision and recall on the Bible pairs against the author's own text, and run twice.
• I2. Report the link rate between words and concepts: the share of source words that reach a concept in the meta-world, per author. This is the ontology-linking statistic that the report gives, and it shows where the lexicon is thin.
• I3. Measure the time an editor needs to correct the output, as the report does for translators. For Mova, use the number of edits on the dev verses as the proxy, with the same edit script for every run.

For Mova

Article insertion is a clean case of open problem 4 (function words as features). The code is tools/translate/src/profile.rs for the encoder rules and meta.rs for the concept links. The rule-learning idea is the part to keep; the report's statistical rules are not described in enough detail to copy, so the counted version has to be built and measured from scratch.

Topics: Machine translation

Cite

The paper itself, not this card. BibTeX from the library's record of the work.

@inproceedings{hovy1994pangloss,
  title = {{PANGLOSS: Knowledge-Based Machine Translation}},
  author = {Hovy, Eduard},
  booktitle = {Human Language Technology: Proceedings of a Workshop held at Plainsboro, New Jersey, March 8-11, 1994},
  year = {1994},
  url = {https://aclanthology.org/H94-1121/}
}

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/hovy-1994-pangloss-knowledge-based-machine-translation/